OpenAI releases GPT-5.2, outperforming Gemini 3 in coding and reasoning benchmarks

Four weeks after GPT-5.1, OpenAI has released GPT-5.2, a new model series designed for professional knowledge work that matches or exceeds Google's Gemini 3 Pro in several key benchmarks.

GPT-5.2 Thinking achieves expert-level performance on the GDPval benchmark (70.9%) and significantly improves coding scores on SWE-Bench Pro (55.6%) and SWE-Verified (80%).
The model demonstrates massive gains in abstract reasoning, reaching 52.9% on ARC-AGI-2 compared to GPT-5.1's 17.6%, while also cutting response error rates by 30%.
Three variants are available: GPT-5.2 Instant, Thinking, and Pro, with API pricing increased to $1.75/$14 per million tokens for the standard tier.

The release aims to address previous competitive pressures from Google by delivering faster, cheaper, and more accurate results for complex tasks like financial modeling and contract analysis.

Benchmarks

Benchmark	Model	Score
τ²-bench	GPT-5.2 Thinking	98.7%
ARC-AGI 1	GPT-5.2 Pro	90.5%
CharXiv	GPT-5.2 Thinking	88.7%
SWE-bench Verified	GPT-5.2 Thinking	80%
SWE-bench Pro	GPT-5.2 Thinking	55.6%
ARC-AGI 2	GPT-5.2 Thinking	52.9%

Benchmark

Model

Score

τ²-bench

GPT-5.2 Thinking

98.7%

ARC-AGI 1

GPT-5.2 Pro

90.5%

CharXiv

GPT-5.2 Thinking

88.7%

SWE-bench Verified

GPT-5.2 Thinking

80%

SWE-bench Pro

GPT-5.2 Thinking

55.6%

ARC-AGI 2

GPT-5.2 Thinking

52.9%